Non-invasive time inference of repetitive contusion injuries using hyperspectral imaging and machine learning
Shilin Miao1, Lele Tian1, Xiangyan Zhang1
1Department of Forensic Science, Xiangya School of Basic Medical Science, Central South University, Changsha 410013, China.
Background:
Objective wound age estimation remains challenging in forensic practice, particularly for repetitive contusion injuries (RCI), where overlapping hemorrhage and repair processes limit the reliability of visual assessment.
Objective:
This study aimed to develop a non-invasive framework to infer post-injury time for RCI using hyperspectral imaging (HSI) and machine learning (ML), with histopathology as biological reference.
Methods:
RCI was induced on the dorsal skin of mice once daily for up to 8 consecutive days. Reflectance hyperspectral data (502.4-899.7 nm) were acquired over an 8-day period and compared with hematoxylin-eosin (HE) staining findings. Spectra were preprocessed using Savitzky-Golay first derivative (SG1), standard normal variate (SNV), or multiplicative scatter correction (MSC). Eight ML classifiers (RF, SVM, XGBoost, DT, KNN, GNB, LDA, and MLPNN) were evaluated using full spectra and principal component analysis (PCA) features (cumulative explained variance >99%).
Results:
The spectra exhibited time-dependent changes with characteristic hemoglobin-related features near 555 nm and 576 nm and a near-infrared feature around 760 nm, consistent with evolving hemorrhage and tissue response observed by HE staining. Among the tested models, SVM achieved the best performance, reaching a testing accuracy of 93.8% with SNV preprocessing; PCA retained comparable performance while reducing dimensionality.
Conclusion:
These findings suggest that hyperspectral imaging, combined with suitable preprocessing and machine learning, may support more objective, non-invasive time inference of repetitive contusion progression in this mouse model, and warrants further validation for translational forensic applications.
